Retinal Blood Vessel Segmentation Algorithm

The purpose of this project is to train autoencoder model that can classify pixels from retinal images into two categories -- blood vessel and background.

Double-Attention, Residual, Efficient U-Net

Model architecture is based on U-Net, which is improved by implementing residual connections and attention modules.

Results

Model performance was evaluated on 16 different databases which vary in terms of resolution, brightness and FOV. Most datasets use narrow-angle images, and 4 datasets are obtained by wide-angle funduscameras. Presented results were achieved on fine tuned models.

Database	Accuracy	Precision	Recall	Specificity	F1-score	IoU
CHASE_DB1	0.9742	0.7826	0.8477	0.9833	0.8133	0.6854
AFIO	0.9613	0.7305	0.7566	0.978	0.7402	0.5897
ORVS	0.976	0.7394	0.795	0.9854	0.7659	0.622
LES-AV	0.979	0.8998	0.8102	0.9927	0.8527	0.7432
STARE	0.9633	0.7195	0.8108	0.9755	0.7612	0.6204
DRIVE	0.9684	0.7923	0.8236	0.9813	0.8068	0.6764
DualModal2019	0.9793	0.8734	0.823	0.991	0.8468	0.7348
IOSTAR	0.9725	0.8259	0.796	0.9863	0.8083	0.6792
TREND	0.9706	0.6599	0.6804	0.9839	0.6677	0.5018
DR~HAGIS	0.9817	0.7733	0.7737	0.9906	0.7729	0.632
HRF	0.9735	0.8364	0.8585	0.9844	0.846	0.734
ARIA	0.9528	0.6803	0.7834	0.9673	0.727	0.5719
VAMPIRE	0.9827	0.7104	0.7964	0.989	0.751	0.6013
RECOVERY	0.9627	0.8046	0.8046	0.9794	0.8046	0.6731
PRIME-FP20	0.9947	0.7832	0.7592	0.9975	0.7664	0.6228
Own dataset	0.9867	0.8934	0.849	0.9944	0.8707	0.771

Application

Models can be trained or used via specialized methods, or by running main.py. Configurations related to image preprocessing, filepaths other hyperparameters can be configured via config.py.

To run application, create folders with datasets in Datasets/ directory. Image division and augmentation will be done automatically and saved in folder declared in config.py.

Program can be run with IDE or via CLI. Supported arguments are listed in utils.py.

Pre-trained models

Repository contains pre-trained models which can be used for inference or further fine-tuning. To run segmentation task you have to adjust config.py to match pre-trained model filepath and set output images directory. Then run test.py. Console will display segmentation metrics and segmentation results with annotated missclafficiations will be saved in specified directory.